DUDA pretrained release

Pretrained assets for Discovering Unseen Degradations to Adapt Open-World Image Restoration (NeurIPS 2026).

• Project: https://bo-oseng.github.io/DUDA/
• Code: https://github.com/bo-oseng/DUDA
• Dataset: https://huggingface.co/datasets/ses7720/DUDA-WRES

Files

duda-pretrained.tar.gz contains the paper assets (about 193 MiB compressed):

Asset	Purpose
checkpoints/stage0_latest.pth	Supervised restoration initialization, epoch 500
checkpoints/semi_supervised_epoch10.pth	Adapted student/teacher and descriptor, epoch 10, step 14,470
cgcd/	PCA, scaler, class mapping and stage Gaussian parameters
dino_adapter/	Learned DINOv3 BOFT adapter, permutation buffers, base-layer biases and classification head
features/wres_with_clear/dinov3-vitl16-sl/class_order.txt	Degradation class ordering
manifest.json	Per-file SHA-256 checksums

The Docker image below includes the runtime and implementation, so a GitHub checkout is optional for Docker users.

DINOv3 base weights are obtained from the official gated model repository separately. The learned BOFT adapter needs its saved boft_P buffers and base_bias.pt.

Download and run

hf download ses7720/DUDA --include "duda-pretrained.tar.gz*" README.md SHA256SUMS release.json --local-dir downloads/duda-models
cd downloads/duda-models
sha256sum -c SHA256SUMS
tar -xzf duda-pretrained.tar.gz

From a DUDA source checkout, follow its README.md to build the Conda environment or Docker image. Obtain DINOv3 using the access/download instructions in README.md, then link the extracted release assets:

python scripts/link_assets.py \
  --features /path/to/duda-pretrained/features \
  --dino-adapter /path/to/duda-pretrained/dino_adapter \
  --cgcd /path/to/duda-pretrained/cgcd \
  --stage0 /path/to/duda-pretrained/checkpoints/stage0_latest.pth \
  --adapted /path/to/duda-pretrained/checkpoints/semi_supervised_epoch10.pth
CUDA_VISIBLE_DEVICES=0 python restoration/infer.py \
  --input /path/to/image.jpg --output outputs/restored.png

The default restoration checkpoint uses the adapted teacher. Add --stage 0 for supervised initialization. Dataset downloads provide two paths: generate patches from the original training images, or download the optional precomputed patch archive. See the dataset card for both commands.

Validation used Python 3.11, PyTorch 2.4.0, Transformers 4.56.0, PEFT 0.12.0, CUDA 12.1 and an A100. The checks cover strict checkpoint loading and inference of one full-resolution image through DINOv3, BOFT, the descriptor and OneRestore using both supervised and adapted checkpoints. They do not constitute full training or a rerun of the complete benchmark.

Continual 6+1+2 and 6+1+1+1 streams

The optional continual/duda-continual-weights.tar.gz contains five regular-stream restoration checkpoints (two stages for 6+1+2 and three for 6+1+1+1), plus each stream's PCA, scaler, class mappings, and stage-specific CGCD means/covariances. The stage-aware source and Bash scripts are maintained in the DUDA GitHub repository and documented in CONTINUAL.md. The Docker image includes these scripts. The AIC variants are not included.

hf download ses7720/DUDA --include 'continual/duda-continual-weights.tar.gz' --local-dir downloads/duda-models
cd /path/to/DUDA
tar -xzf /path/to/downloads/duda-models/continual/duda-continual-weights.tar.gz --strip-components=1
STAGE=2 bash bash/test_6p1p2.sh
STAGE=3 bash bash/test_6p1p1p1.sh

The two final stages were checked on one image with strict checkpoint loading and DINOv3/BOFT inference. Full Table 3 metrics were not rerun for this add-on.

Downloadable Docker image

The optional docker/duda-neurips2026.tar.zst file (7.00 GB compressed) contains the tested Docker image, including the runtime and project code. It is separate from the model bundle. Model weights, DINOv3 base weights and datasets are obtained separately.

Download and load the image:

hf download ses7720/DUDA --include "docker/*" --local-dir downloads/duda-docker
(cd downloads/duda-docker/docker && sha256sum -c SHA256SUMS)
zstd -dc downloads/duda-docker/docker/duda-neurips2026.tar.zst | docker load

The loaded tag is duda:neurips2026. This is an image archive for docker load; it is not a Docker registry endpoint.

To run one image, first extract duda-pretrained.tar.gz in your working directory. Request access to the official DINOv3 model, then download its base weights:

hf auth login
hf download facebook/dinov3-vitl16-pretrain-lvd1689m \
  --revision ea8dc2863c51be0a264bab82070e3e8836b02d51 \
  --include config.json model.safetensors preprocessor_config.json LICENSE.md README.md \
  --local-dir dinov3-vitl16

Set your input image path and run the container. Change the host GPU selection as needed; the process inside the container uses the single exposed GPU.

export DUDA_INPUT=/absolute/path/to/input.jpg
mkdir -p outputs
docker run --rm --gpus '"device=0"' --shm-size=2g \
  -v "$PWD/dinov3-vitl16:/app/assets/dino_base:ro" \
  -v "$PWD/duda-pretrained/dino_adapter:/app/assets/dino_adapter:ro" \
  -v "$PWD/duda-pretrained/cgcd:/app/assets/cgcd:ro" \
  -v "$PWD/duda-pretrained/features:/app/assets/features:ro" \
  -v "$PWD/duda-pretrained/checkpoints:/app/assets/checkpoints:ro" \
  -v "$DUDA_INPUT:/inputs/input.jpg:ro" \
  -v "$PWD/outputs:/app/outputs" \
  duda:neurips2026 python restoration/infer.py \
    --input /inputs/input.jpg --output outputs/restored.png

The image pins Pillow 11.3.0, whose libjpeg-turbo JPEG decoding matches the paper runs; with this image, the evaluation output of the main adapted checkpoint is pixel-identical to the original per-image record on the checked image.

The image can also run scripts/prepare_datasets.py for dataset patch generation and monitoring-subset preparation; see the dataset card.